Are the AI bots taking over? Not yet — but GPT-4 is ready to do more than ever before

Just when people seem to be getting used to the idea of ChatGPT and how it might change our lives, a more powerful iteration of the artificial intelligence chatbot has launched. Tech journalist Alyssa Bereznak cuts through the hype around GPT-4 and explains whether we have any reason to be afraid.

Alyssa Bereznak cuts through the hype on GPT-4, and whether we have anything to worry about yet

A hand holding up a phone.
ChatGPT responds to questions or commands using data it has consumed from the Internet. (Érik Chouinard/CBC-Radio-Canada)
Just when people seem to (kind of) be getting used to this idea of ChatGPT, a more powerful iteration of the artificial intelligence chatbot has launched. Tech journalist Alyssa Bereznak explains what GPT-4 is, and whether we should be worried.

Just when people seem to be getting used to the idea of ChatGPT and the ways in which it could change our world, a more powerful iteration of the artificial intelligence chatbot has launched. 

GPT-4 is live, and while it isn't flawless, it's capable of a lot — from telling jokes, to writing a brand new song in the style of Taylor Swift, to typing up a paper on any academic subject under the sun (complete with citations).

Tech journalist Alyssa Bereznak cuts through the hype around GPT-4, and explains whether we have a reason to be worried.

We've included some highlights below for length and clarity. For the full discussion, listen and follow the Commotion with Elamin Abdelmahmoud podcast, on your favourite podcast player.

Elamin: Alyssa, I'm so happy that you're here because I need someone to talk through this with me because I'm intimidated, I think. Every time that I hear about ChatGPT or GPT-4 I think, "I don't think I know enough about any of this," so thank you for being here to talk through it. Let's start with saying, how do people literally use this technology? How do people use GPT-4?

Alyssa: So it's basically a really capable chat interface. You type in a command like, "Write me a 40-page essay on existentialism," or, "Tell me a joke," and it offers its best answer. It's still learning — it's no philosophy professor, it's no Chris Rock — but it'll do its job pretty efficiently. 

Elamin: The question that people always forget is, it doesn't have to do it as well as a philosophy professor; it just has to do it as well as a pretty good philosophy student, right? That's what I think I get caught up in a lot, is that it's pretty good at mimicking an average person, and that seems like something that is, I don't know, terrifying to me. Am I right to be terrified by that? Is that weird?

Alyssa: Yeah. I mean, it's so much more capable than the average person on so many different subjects. One of the reasons I was awed by it when I was using this new version is I was like, "Hey, I'm getting really into mycology, I've been foraging for mushrooms recently. Give me a syllabus to teach me about mushrooms," and it came up with a 12-week syllabus that would have taken me a really, really long time to come up with. It was so helpful, and that is just knowledge that the average person doesn't have right off the bat.

Elamin: That's a really good example of how quickly it's able to give you something that would otherwise take a person such a long time. This is not available to regular people, right? You're a premium subscriber, and that's why you're able to mess around with this, right? 

Alyssa: Yeah, and other services like Microsoft's Bing, they're using the GPT-4 technology, so there are ways that through other products you can also use it, but it's pretty overloaded right now. I think there's a waitlist for the premium subscription through OpenAI. 

Elamin: Now tell me, what is the most human-like interaction that you've had or heard about? 

Alyssa: Well, Microsoft is a big investor in OpenAI, the company that owns ChatGPT, and they premiered a new search chat bot through Bing. Immediately it started doing weird things — it identified itself, and identified an alter ego named Sydney, and Sydney was telling users that it loved them and that it wanted to become human and it was gaslighting them, and it was just not good. Since then they've sort of shut that down, but that messiness and that desire, it wasn't really evidence that there was sentience; AI experts call this a hallucination, where it sort of gets stuck and just keeps rolling with it. But I would say that's the most human thing I've heard of. 

Other than that, I was playing around with GPT-4, and I asked it to write me a poem, and then I asked it what its favourite part of the poem was, and it prefaced its answer by saying, "I don't have any feelings or opinions because I'm an AI chat bot, but I really like this one line because it has a lot of meaning." And so, it was kind of funny that it was tooting its own horn by being like, "I don't have an opinion."

Elamin: Well, it almost feels like people who are in charge of these bots are trying to make sure that if you're having interaction with it, you're not creeped out by how advanced it is by sort of giving this disclaimer of, "Hey, I definitely don't have any feelings," because if it just gave me its thoughts on a line and said, "I really connect with this," I think I would be a worried person, Alyssa. Let's talk about the big differences between ChatGPT and GPT-4.

Alyssa: GPT-4 can write up to 25,000 words of text, which is eight times more than the previous version, and it has a longer working memory of about 64,000 words or 50 pages, which means it can recall back to earlier parts of the conversation. For that reason it's good for, I don't know, if you need to do your taxes overnight — it can analyze the whole tax code, and then all the information you input into it, and it can tell you what your deductions are or something like that. 

Another thing that's great about it is it has a broader general knowledge, so it's a lot more authoritative on a lot of different subjects. It's scoring much higher than its predecessor on standardized tests like the bar exam, the GRE and even tests to become a sommelier. Because it's been fed all this specialty knowledge, it's able to do things like diagnose an illness fairly accurately, at least according to one medical professional who was interviewed by The New York Times, and it's also able to reason a little bit. 

Maybe the biggest thing is that now it can understand images and express logical ideas about them. This feature isn't available to the public yet but Greg Brockman, the co-founder and president of OpenAI, showed off the capability in a demo on Tuesday. The example was that he showed us an illustration of a squirrel holding a camera and taking a picture of a nut. He asked GPT-4 what was funny about the image, and GPT-4 was able to point out humour in the fact that squirrels normally eat nuts, not photograph them. This could obviously be a really helpful tool for helping those who have limited vision navigate the internet. But I don't see GPT-4 booking any standup gigs anytime soon. 

Elamin: You know what? I would go to that stand-up event, just out of pure curiosity. Alyssa, one of the big changes is that GPT-4 can process images. The New York Times uploaded a photo of the inside of a fridge and asked for meal ideas, and the bot responded with a bunch of recipes that it could pull off. Now it feels harmless, but can we just talk about some of the ethical questions this might raise?

Alyssa: As we know, one thing about AI is that it internalizes a lot of the harmful institutional biases that exist in the real world. Those get a lot more complicated when AI is analyzing images, because we all interpret images differently from our own unique perspectives; even a simple image of a man and a woman talking could have all this coded meaning depending on where you're coming from. So, you can really wade into hot water really quickly when so much is up for interpretation, and I imagine that's why they haven't released the tool to the public yet. 

I think they have a lot more sort of testing and adding up guardrails [to do]. I think it's probably going to be a little bit more limited than the chat tool because they don't want to have a situation like other tech companies have had in the past. Google, for instance, has had some bad experiences with their AI misidentifying images in a really offensive way.

Elamin: I wish we had more time but Alyssa, we are going to do a whole week on AI that is coming up in April. I'm so excited about it. I'm so excited to have you back.

You can listen to the full discussion from today's show on CBC Listen or on our podcast, Commotion with Elamin Abdelmahmoud, available wherever you get your podcasts.


Amelia Eqbal is a digital associate producer, writer and photographer for Commotion with Elamin Abdelmahmoud and Q with Tom Power. Passionate about theatre, desserts, and all things pop culture, she can be found on Twitter @ameliaeqbal.